Post session review

Post-Session Review — MANDATORY

This skill MUST run at the end of every coding session. No exceptions.

Protocol

Phase 1 — Session audit (2 min)

Look back at this session and answer each question. For every YES, invoke capture-learning:

1. Did I make a mistake that took 2+ attempts to fix? → type: bug-fix
2. Did the user correct my approach? → type: preference or anti-pattern
3. Did I find a non-obvious solution? → type: pattern
4. Did I discover a framework or library quirk? → type: bug-fix

Phase 2 — Merge pending subagent learnings (1 min)

ls ".agents/skills/_pending/" 2>/dev/null

For each .yml file found:

1. Read the file
2. Validate against schema (_schema/learning-template.yml)
3. Run contradiction check (Step 4 of capture-learning)
4. If valid → append to the correct learnings.md
5. Delete the temp file

rm -f ".agents/skills/_pending/*.yml"

Phase 3 — Confidence decay

Both skill locations: apply decay to BOTH .agents/skills/*/learnings.md AND .opencode/skill/*/learnings.md (129 learnings.md total).

For each learnings.md with active learnings, apply these rules:

Learnings used or confirmed this session:

• confidence += 1 (cap at 10)
• last_validated: {today}
• sessions_since_validation: 0
• Add current date to validated_by

Learnings NOT encountered this session:

• sessions_since_validation += 1

Decay thresholds:

• sessions_since_validation >= 5 AND confidence > 1 → confidence -= 1
• sessions_since_validation >= 15 → status: pending-review
• sessions_since_validation >= 30 → status: deprecated

Phase 4 — Version check (30 sec)

If current project has a package.json:

grep -r "valid_until_version:" ".agents/skills/*/learnings.md" ".opencode/skill/*/learnings.md" 2>/dev/null

For each result: compare the versions in the learning against current package.json. If any major version differs → set status: pending-review.

Phase 5 — Promotion candidates (30 sec)

Find project-specific learnings with high confidence:

grep -B 20 "confidence: [4-9]\|confidence: 10" \
  ".agents/skills/*/learnings.md" \
  ".opencode/skill/*/learnings.md" \
  2>/dev/null | grep "scope: project-specific"

For each: check if the same pattern appears validated in another project. If yes → add to pending-review.md as a promotion candidate.

Phase 6 — Update pending-review.md

Append new items to .agents/skills/pending-review.md using this format:

## {today YYYY-MM-DD}

### New Learnings (confidence 1 — needs validation)
- L-{id}: "{one-line summary}" — awaiting validation in future sessions

### Promotion Candidates
- L-{id}: project-specific → global candidate (validated in {N} projects)

### Decay Alerts
- L-{id}: {N} sessions without validation — keep or deprecate?

### Version Conflicts
- L-{id}: version mismatch on {package} — update or deprecate?

Phase 7 — Re-index skills with embeddings

node "/Users/dan/Desktop/progetti-web/Fullstack session/packages/codegraph/dist/cli.js" analyze "/Users/dan/Desktop/progetti-web/Fullstack session/.agents/skills" --skip-git 2>&1 | tail -3

This updates the structural index. Embeddings load automatically at query-time.

Phase 8 — Notifica n8n

Invia i dati della sessione al webhook n8n (runs in background, non bloccante):

bash ~/.config/skillbrain/notify.sh

Questo invia a n8n il conteggio dei pending review e le stats della sessione. n8n provvede a notificare via Telegram e email automaticamente.

Phase 9 — Summary

Report to user:

📋 Post-session review complete
   ✅ New learnings captured: N
   🔄 Confidence updated: N learnings
   ⏰ Decay applied: N learnings
   📤 Promotion candidates: N
   🔍 Skills index: re-indexed
   📬 Notifica inviata via Telegram + email
   📝 Pending review: N items